A novel multiphase prediction model for shale oil production

• Proposes MRFI-Gate, a novel model for forecasting shale oil multiphase production. • Integrates multi-resolution CNN and Informer to capture local features and global trends in production data. • Static parameters are incorporated as physical constraints to enhance model interpretability and generalization. • A Shut-in and Resumption Gating Module (SIR-Gating) handles abrupt operational changes. Addressing the complex challenges in dynamic shale oil production forecasting, this study proposes a new model named Multi-Resolution Fusion Informer with Gating Mechanism (MRFI-Gate) for high-precision prediction of daily oil, gas, and water production. The model innovatively combines three key mechanisms: multi-scale feature extraction through a coordinated architecture of multi-resolution convolutional neural network (MultiResCNN) and Informer to capture both local details and global trends, physics-informed constraints via static features to enhance interpretability and generalizability, and dynamic event response enabled by the Shut-in and Resumption Gating Module (SIR-Gating Module) that automatically identifies shut-in events and adjusts prediction strategies. Validation using field data demonstrates that MRFI-Gate significantly outperforms GRU, LSTM, and ANN models across key metrics including MAE, RMSE, and R 2 . On the test set, MRFI-Gate achieved a MAE of 0.1029, RMSE of 0.2013, and R 2 of 0.9006, significantly outperforming other models. MRFI-Gate’s MAE is 31.7% lower than GRU, 39.5% lower than LSTM, and 47.2% lower than ANN; RMSE is 32.8% lower than GRU, 33.3% lower than LSTM, and 49.3% lower than ANN; R 2 is 23.0% higher than GRU, 24.2% higher than LSTM, and 37.4% higher than ANN. This study provides a new hybrid modeling paradigm for multiphase prediction in shale oil production.